GPT-SoVITS

GPT-SoVITS pairs a GPT-style semantic model with SoVITS voice synthesis for few-shot voice cloning and text-to-speech. Given only about five seconds of reference audio it performs zero-shot cloning, and with roughly one minute of data it can be fine-tuned for markedly better speaker similarity; it also does voice conversion between speakers. The system covers Chinese, English, Japanese, Korean, and Cantonese and can speak languages different from the reference voice's training data. A WebUI bundles the whole workflow, including tools for vocal separation, audio slicing, ASR-based labeling, and dataset preparation, so non-experts can train a voice end to end. Pretrained on around 5,000 hours of speech, it runs fastest on NVIDIA GPUs, with real-time factors near 0.03 on an RTX 4060 Ti, but also supports CPU and Apple Silicon. Released under the MIT license, it is a common choice for character voices, dubbing, and hobbyist voice cloning.